Componentix blog

In one of our Grails applications we had to run a number of batch jobs. Nothing unusual and Grails supports it quite well with the excellent Quartz plugin.

But when we deployed application in production, we noticed that after running for some time, it consumed a lot of memory and JVM was spending all the time running garbage collection. The reason for it was that our jobs were quite long-running, taking several hours to complete, and Grails wasn’t really designed for such kind of use case.

First problem is actually well-known and documented — using single Hibernate session for a long time gets a lot of objects being added to cache. It is described for example in this blog post.
This basically goes into flushing and clearing Hibernate session like following:

sessionFactory.currentSession.flush() 
sessionFactory.currentSession.clear()

But while less memory was being leaked after that, the problem still wasn’t resolved. However thankfully to this nice post about Grails memory leak I learned about another more obscure problem and got it fixed.

Main idea is that save() method stores validation errors info in thread-local storage when HttpServletRequest instance is not available. So this info needs to be cleared from time to time, which can be achieved by something like this:

DomainClassGrailsPlugin.PROPERTY_INSTANCE_MAP.get().clear()

Resolving these two issues solved our problem with memory usage. Hopefully this post prevents you from spending much time resolving the same issues.

In one of our Grails projects we had to use Microsoft SQL Server as a database. Hibernate has support for it and works good enough, however the schema generated by it is not ideal, at least from our point of view. In particular, properties of boolean type were stored in columns of type INT, while both SQL Server 2000 and 2005 support special type just for booleans — BIT; all Ids, which usually have type long, had to be squeezed into columns of type INT (again this stupid integer!); such SQL types as TEXT and IMAGE were used, instead of (preferred for SQL Server) VARCHAR(max) and VARBINARY(max) respectively; and finally (for me — the most annoying thing) strings were stored in non-Unicode sequences. The very last point is also very unhandy, since strings in Java are stored in Unicode, and having to deal with encodings explicitly when storing or reading strings or entire texts from the database is a bit against the simplicity we expect from Hibernate when we use it.

So I decided to develop customized Hibernate dialect for MS SQL Server. To do this, I started digging the source code of Hibernate (the dialect classes to be precise), and soon I found out that Hibernate’s SQLServerDialect extends from SybaseDialect, which is responsible for most of the “sins” outlined in the paragraph above. Don’t know why they haven’t overridden any of that in a subclass, probably they had a good reason not to do so, but for our company this reasoning was certainly not working out. Just extending the standard implementation was fairly enough to solve the problem, actually it was sufficient just to register appropriate column types in the constructor. The end result looks like this:

public class SQLServerDialect extends org.hibernate.dialect.SQLServerDialect {
 
   /**
    * Initializes a new instance of the {@link SQLServerDialect} class.
    */
    public SQLServerDialect() {
        registerColumnType(Types.BIGINT, "bigint");
        registerColumnType(Types.BIT, "bit");
        registerColumnType(Types.CHAR, "nchar(1)");
        registerColumnType(Types.VARCHAR, 4000, "nvarchar($l)");
        registerColumnType(Types.VARCHAR, "nvarchar(max)");
        registerColumnType(Types.VARBINARY, 4000, "varbinary($1)");
        registerColumnType(Types.VARBINARY, "varbinary(max)");
        registerColumnType(Types.BLOB, "varbinary(max)");
        registerColumnType(Types.CLOB, "nvarchar(max)");
    }
}

Obviously, similar technique can be used with any custom Hibernate dialect, not just for MS SQL. Feel free to use it with any other DB engine, etc. If you find it useful, you can fork our code snippet at Github: http://gist.github.com/225543